Supplier invoice reconciliation is the quiet bottleneck of every wholesaler: before goods can be received into stock, someone has to check the supplier’s invoice line by line against the purchase order in the ERP. For a distributor of fastening and mounting hardware with around 35,000 items and several branches, that was a daily job done with two windows open side by side. We built an AI reconciliation assistant that pairs the lines, checks quantities and prices, and hands a person only the lines that actually need a decision.
The problem: nothing ever matches directly
Comparing two documents sounds simple. In practice, almost nothing lines up one to one:
- Three identities per product. The company’s own item code, the supplier’s code and the barcode, and often only one of them appears on the invoice.
- Different units. The supplier invoices in boxes; the warehouse counts pieces.
- Different price bases. Some suppliers quote a price per 100 or per 1,000 pieces.
- Different order. The lines on the invoice follow the supplier’s logic, not the purchase order.
- Missing references. The invoice often does not carry the buyer’s order number, and freight and pallets appear as extra lines nobody ordered.
So the work is not the comparison itself. It is the translation that has to happen before any comparison can start, repeated for every document, every day.
What we built
An assistant that runs next to the ERP and works through each supplier case in a few steps:
- It collects the documents. Suppliers’ order confirmations and invoices arrive in a dedicated mailbox. The assistant recognises the document type (invoice, order confirmation, packing list, quotation) and sets aside what is not needed.
- It reads the order from the ERP. The purchase order is fetched by its number, so nobody exports spreadsheets by hand.
- It pairs the lines by product, not by position. Matching uses the internal code, supplier code, barcode or name, converts boxes to pieces and “per 100” prices to unit prices, and marks freight and pallets as non-product lines.
- It sets a confidence level for every case. Green: everything agrees, one button to approve. Yellow: only the lines with a quantity or price difference, with a suggested resolution. Red: unknown supplier, totals that do not add up or a duplicate go to a “needs decision” queue and the responsible manager is notified.
- It prepares the result for the ERP. After approval, a purchase draft with the right quantities, prices, invoice number and dates is created in the ERP, together with a list of discrepancies to take back to the supplier.
The assistant learns from corrections. When someone confirms once that a supplier code is a particular internal item, or that a box holds 200 pieces, it is saved as a rule and applied next time.
What changes for the team
| Before | With the reconciliation assistant | |
|---|---|---|
| Starting the work | Export the order, open the invoice, compare in two windows | Documents are picked up from the mailbox |
| Matching lines | By eye, translating codes, units and price bases | Paired automatically by product identity |
| What a person reviews | Every line of every invoice | Only yellow and red lines, with a suggested fix |
| Discrepancies | Noted by hand, easy to lose | Listed per case, ready to send back to the supplier |
| Processing time per invoice | Depends on the size of the order | Seconds of processing, then the human review |
| Knowledge about suppliers | In one or two people’s heads | Stored as approved rules, visible in an audit log |
The system runs on a modest virtual machine on the company’s own server and is sized for 50 to 150 documents a day on 1 CPU core and 1 GB of memory. Documents never leave the company network.
Where people stay in control
Nothing is posted automatically. A person approves every case before a draft is created in the ERP, and the final posting in the ERP stays with the warehouse team, exactly as before. Every rule the assistant learns is approved by a person and recorded in the audit log.
We were also clear about the limit: the first release works with PDFs that contain a text layer and with Excel files, which by the team’s own estimate covers roughly nine documents out of ten. Photos and scans are escalated to a person rather than guessed.
Where else this works
The same reconciliation logic fits anywhere two documents have to agree before money moves:
- Construction: subcontractor bids compared line by line before a contract is awarded.
- Public tenders: conditions, declarations and specifications checked against each other before a bid goes in.
- Financial services: a customer’s company documents checked against onboarding rules.
Related use cases: Subcontractor bid comparison · KYB document checks · Freight quote automation. All Document checks & reconciliation use cases · How we deliver this: Custom AI solutions
FAQ
How does AI match invoice lines when the codes differ?
It matches by product identity rather than by line position, using the internal code, the supplier’s code, the barcode or the product name, and converts units and price bases before comparing. Confirmed matches are saved as rules for next time.
Does the system post invoices into the ERP by itself?
No. After a person approves a case, the assistant creates an unposted purchase draft. Final posting stays with the warehouse team.
Can it run without sending documents to the cloud?
Yes. In this case it runs on the company’s own server next to the ERP, and documents stay inside the company network.
Want to see how many of your invoice lines could be checked automatically? Talk to us